This study presents a graph-based predictive framework for student performance, leveraging Graph Neural Networks (GNNs) to model the complex relationships between student profiles and academic subjects. Students and courses are represented as nodes in a bipartite graph, with enrollment links serving as edges. By learning edge embeddings, the model captures latent interactions often missed by conventional approaches. These representations are then used to train a regression model that estimates student grades across diverse academic faculties. Empirical evaluations show strong predictive accuracy across most datasets, though performance varies with institutional and curricular heterogeneity. These findings underscore the need for adaptable models tailored to specific educational environments. The proposed approach highlights the potential of GNNs for uncovering structural patterns in academic data and supporting personalized learning strategies through data-driven decision-making.